| Summary: | In order to provide better monitoring performance, between-phase nonlinear correlations and differences should be considered and separately monitored in multiphase batch processes. However, it is a more challenging task for batch processes to extract the nonlinear correlation information because the large-size samples may bring the computational complexity and high-dimensional kernel instability. To address the above issue, a feature vector selection with kernel vector correlation analysis (FVS-KVCA) method is developed for batch processes. In this paper, a novel two-level phase division method is firstly proposed to divide batch processes into the steady phases and the transitions. Then, the local phase models are constructed based on the FVS-KVCA method to separate process information into the common information and the specific information, which represent the nonlinear correlation between two neighboring phases and within only one phase, respectively. Based on such an information separation, the transition can be further divided into the transitional phases and the mixing phases by the second-level division. Also, a dynamic transition modeling method is introduced to solve the transition monitoring problem. Finally, online process monitoring can be conducted based on the defined score features to select the right model. The proposed algorithm is applied to the penicillin fermentation process to illustrate the effectiveness and feasibility.
|
|---|
